Research On Active Learning Algorithms Of Bioelectrical Signals

In recent years,with the continuous development of computer technology,the task of automatic identification of bioelectric signals has caused a large number of scholars to conduct research.The main application scenario of automatic identification of bioelectric signals is to assist doctors in diagnosis and treatment.The purpose is to enable computers to autonomously and efficiently judge human diseases and abnormalities reflected by bioelectric signals.However,in practical applications,the task of data labeling of bioelectrical signals has become a big problem.Since the bioelectric signal itself is not easy to be understood by ordinary people,it requires very experienced medical experts to spend a lot of time for labeling,which greatly wastes human resources and time resources,and makes the labeling acquisition cost of the data set very huge.In the computer field,the common method to reduce the amount of annotations and reduce the cost of annotations is active learning.However,after investigation,the application of active learning methods in the field of bioelectric signal classification is close to zero.Only scholars have conducted confirmatory experiments on EEG signals and ECG signals.So far,no scholars have conducted active studies on EMG signals.In the learning experiment,no scholar has used the most advanced active learning method to study the classification of bioelectric signals.In order to solve the problem of labeling cost of bioelectric signals,this paper applies the active learning method to the classification task of bioelectric signals,and proposes corresponding solutions to the problems existing in the active learning method.The main work of this paper is as follows:1.This article uses a variety of active learning methods to carry out a series of experiments on the classification of bioelectric signals.The traditional active learning methods and the most advanced active learning methods are applied to the classification task of bioelectric signals.And an EEG signal data set to verify the effectiveness of active learning.2.Aiming at the problem of falling into the trap of active learning localization due to the existence of unknown unknowns in active learning for bioelectric signals,this paper proposes an active learning method based on query unknown unknowns using subspace projection.This method not only conducts extensive exploration in space,but also conducts exploration in specific areas,thereby effectively discovering unknown unknowns during the iterative process of active learning,reducing the impact of falling into the localization trap of active learning,and improving active learning The effect of reducing the amount of data annotations.3.Aiming at the problem of category specificity in active learning for bioelectrical signals,this paper takes the global category prediction information to weight the information entropy as the main idea,and proposes an active learning method based on classes information supervision.This method encourages active learning to choose data from the class that are difficult to learn for labeling.The method solved this problem,improving the effect of active learning,and reducing the amount of data labeling.4.In order to make the information provided by the data screened by the above two active learning methods complementary,this paper combines the above two active learning methods using a dynamic weighting method.And finally proved in the experiment that the fusion method can bring complementarity to the two active learning methods which enhancing the effect of active learning.